def __init__(self, config, subset):
super(Davis2017OneshotDataset, self).__init__(config, NUM_CLASSES, VOID_LABEL, subset,
image_size=DAVIS2017_IMAGE_SIZE)
self._video = None
data_dir = config.unicode("davis_data_dir", DAVIS2017_DEFAULT_PATH)
self.split = config.unicode("split", "val")
assert self.split in ("val", "dev", "eval")
if self.split == "val":
list_file = "ImageSets/2017/val.txt"
elif self.split == "eval":
list_file = "ImageSets/2017/test-challenge.txt"
else:
list_file = "ImageSets/2017/test-dev.txt"
if not config.bool("adjustable_output_layer", False):
list_file = list_file.replace(".txt", "_ids.txt")
imgs, ans = read_image_and_annotation_list_2017(data_dir + list_file, data_dir)
self._video_tags = unique_list([im.split("/")[-2] for im in imgs])
self.imgs_seqs = group_into_sequences(imgs)
self.ans_seqs = group_into_sequences(ans)
only_first_frame_annotation_available = self.split in ("dev", "eval")
if only_first_frame_annotation_available:
self.ans_seqs = [[ans_seq[0]] + ([None] * (len(ans_seq) - 1)) for ans_seq in self.ans_seqs]
self.use_lucid_data = config.int("lucid_interval", -1) != -1
self.lucid_data_dir = config.unicode("davis_lucid_data_dir", DAVIS2017_LUCID_DEFAULT_PATH)
self.lucid_data_video = None
self.lucid_data_video_current = None
def load_videos(self, fn, data_dir, video_range=None):
load_adaptation_data = self.adaptation_model != ""
if load_adaptation_data:
assert self.config.unicode(
"task") == "offline", self.config.unicode("task")
elif DavisOneshotDataset._video_data is not None:
#use this cache only if not doing offline adaptation!
return DavisOneshotDataset._video_data
print >> log.v4, "loading davis dataset..."
imgs, ans, flows = read_image_and_annotation_list(
fn, data_dir, flow_flag=self.twostream)
video_tags = unique_list([im.split("/")[-2] for im in imgs])
imgs_seqs = group_into_sequences(imgs)
ans_seqs = group_into_sequences(ans)
if load_adaptation_data and self.subset == "train":
#change annotations from groundtruth to adaptation data
for video_tag, ans in zip(video_tags, ans_seqs):
for idx in xrange(1, len(ans)):
ans[idx] = "forwarded/" + self.adaptation_model + "/valid/" + video_tag + (
"/adaptation_%05d.png" % idx)
start = time.time()
videos = [None] * len(imgs_seqs)
if video_range is None:
video_range = [0, len(imgs_seqs)]
# from joblib import Parallel, delayed
# videos[video_range[0]:video_range[1]] = Parallel(n_jobs=20, backend="threading")(delayed(_load_video)(
# imgs, ans, self.flow_dir, self.flow_into_past, self.flow_into_future, self.flow_as_angle) for
# (imgs, ans) in zip(imgs_seqs, ans_seqs)[video_range[0]:video_range[1]])
if self.twostream:
flows_seqs = group_into_sequences(flows)
videos[video_range[0]:video_range[1]] = [
_load_video(imgs, ans, self.flow_dir, self.flow_into_past,
self.flow_into_future, self.flow_as_angle, flows)
for (imgs, ans, flows) in zip(imgs_seqs, ans_seqs, flows_seqs)
[video_range[0]:video_range[1]]
]
else:
videos[video_range[0]:video_range[1]] = [
_load_video(imgs, ans, self.flow_dir, self.flow_into_past,
self.flow_into_future, self.flow_as_angle)
for (imgs, ans) in zip(imgs_seqs, ans_seqs)
[video_range[0]:video_range[1]]
]
DavisOneshotDataset._video_data = (video_tags, videos)
end = time.time()
elapsed = end - start
print >> log.v4, "loaded davis in", elapsed, "seconds"
return DavisOneshotDataset._video_data
def group_into_sequences(fns):
seqs = unique_list([fn.split("/")[-2] for fn in fns])
seq_fns = [[fn for fn in fns if fn.split("/")[-2] == seq] for seq in seqs]
return seq_fns